Examining the hidden labor economy behind AI - YaleNews

Concerns about artificial intelligence (AI) — its regulation, energy consumption, and production of misinformation — are under much active public debate.

Yet one area of concern has received relatively little attention: the treatment of the workers who generate, annotate, and evaluate the massive amounts of data that drive AI.

In his new book, “Platform Extractivism: Data Work and the People Powering Artificial Intelligence” (University of California Press), Julián Posada examines the conditions under which these workers were employed by major digital platform companies during the COVID-19 pandemic.

The major AI developers outsource the necessary data production to these companies, which, Posada found, routinely seek out workers in the world’s cheapest, and sometimes desperate, labor markets. Essentially, he argues, they extract value from these markets by exploiting economic instability.

The focus of much of his research is on Venezuela, which from 2018-2023 ranked in the top five countries for data production work. In a survey he distributed to platform workers there, they self-reported average wages of $1.73 per hour.

Julian Posada reads an excerpt from ‘Platform Extractivism’

Julian Posada reads an excerpt from ‘Platform Extractivism’

“This book argues that the AI industry is enmeshed in a form of extractivism: a transnational system of accumulation, labor exploitation, and power concentration that commodifies knowledge and labor through digital platforms,” Posada, an assistant professor of American Studies, in Yale’s Faculty of Arts and Sciences, writes. “This process depends on the centralization of power through platforms that mediate between workers and AI developers. As critical infrastructures, these platforms structure work arrangements for data production while perpetuating a system of exploitation that accumulates wealth for a minority.”

Posada is also founding co-director of the Certificate in Computing, Culture & Society. He is currently teaching a course on “Platforms and Cultural Production,” which examines the cultural and economic impacts of major platforms like Instagram, YouTube, Uber, and Amazon.

Posada sat down with Yale News to talk about the connection between platform extractivism and colonialism, why Venezuela became an attractive labor pool for these platforms, and what can be done to improve conditions for data workers. The conversation has been condensed and edited.

Would you describe what these platform companies do?

Julián Posada: AI has been around for 70 years. Machine learning also has a long history, but deep learning took off more recently, in the 2010s. We have a lot of data around us because we are connected to the internet, but for the most part, it’s unstructured data. It’s data that is just harvested from your phone, your interaction with your email, from talking to Alexa, and so on. For many AI applications, companies need people to label or classify training data, create examples, and evaluate model outputs. For example, by giving feedback to their models, basically telling the AI if it’s doing a good job or a bad job. For this, AI companies often go to an outsourcing firm to reduce costs. And those firms frequently recruit workers through digital platforms in countries where they can pay people much less, although more recently they have also started to recruit workers in higher-income countries, particularly for specialized tasks.

A key point you make in the book is that platform extractivism is entangled with legacies of capitalism and colonialism. Would you talk about that?

Posada: Yes, one of the claims of the book is that the phenomenon I’m observing is not reinventing the wheel. What we are seeing here is a form of outsourcing and extraction of resources. Only in this case, it’s a resource called data. Yes, digital platforms are more recent, AI is more recent, but the way we’re making this AI depends on these older legacies, older patterns, systems that have existed long before.

The Venezuelan case is interesting because the economy had long been heavily dependent on oil revenues. That dependence made it vulnerable to oil-price shocks and tied development to a finite resource. But getting commodities like oil, rubber, and minerals from one place and bringing them to another to create value is a pattern that is centuries old. Dependency theory investigates this. It examines how unequal relationships between economies can channel wealth outward and constrain development in territories supplying resources and labor. What I do in the book is say that they’re doing the same thing with AI data.

How did Venezuela become a target for these companies?

Posada: Venezuela became very prominent in the late 2010s for two reasons. One of them was the economic crisis, which made labor very cheap. Hyperinflation devastated the local economy, so receiving small payments in dollars became attractive to people with fewer income opportunities. Second, people had access to computers and subsidized internet and electricity, supported by earlier government investment. That infrastructure made online work possible, although outages and unreliable connections complicated it.

In researching and interviewing data workers in Venezuela, you found this pattern that you call “up, down, plateau” in terms of the patterns of compensation for workers. Would you explain?

Posada: It also comes from what I learned from studying platform theory from business and economics. Some platform companies spend heavily on incentives when entering a market to attract people to the platform. Say it’s Uber coming to your city and it promises to give you $20 off your first 10 rides. By doing that, Uber gets people hooked on its platform. But then, as the company seeks to reduce costs, it may start reducing or removing those bonuses. The idea is to build a large user base. Over time [in the case of data workers], workers become dependent on the income they earn through the platform. Once you create dependence, then you…can set the rules. In my research, I found that some platforms attracted workers with initially higher pay or bonuses and then reduced those incentives once the workers became dependent on the income. Up, down, plateau.

Another thing that I found is the value being generated for these companies comes from the workers’ networks — the family, friends, colleagues, who are helping to support the worker who is living on such low wages. It’s an extraction from the workers’ territories, not just themselves as individuals.

These workers largely work on their own, at home. What effect does this have on their ability to advocate for higher wages?

Posada: There are two things here: being invisible and being distributed. Let’s start with distributed. One of the big differences between this platform and, say, DoorDash or Uber, is that those workers can see each other when they’re on the job. When they see each other, they can talk and even start organizing, unionizing. In the case of data work, the workers are in their separate houses. They’re not going to meet each other in person, although they could meet online.

The second thing is being invisible. Workers can be managed through account identifiers and performance metrics, even though the company also holds some of their personal information. But for the most part, the platform sees these workers as randomized account numbers. It’s much more complicated than, say, during the industrial revolution, when you had all the factory workers toiling together, they knew each other for years, they had strong bonds. Here, they are distributed. They might know each other from social media, but it’s a very different type of interaction, and that makes it more complicated for them to gain the upper hand in negotiations, or anything related to improving their work situation.

The companies evaluate the workers using algorithmic management, which rates them on metrics like speed and accuracy. If workers question the validity or fairness of those metrics, they are at risk of being fired, yes?

Posada: Yes. Some platforms’ contracts give them broad discretion to deactivate accounts without prior notice and ban workers from returning to the platform. And they reserve the right to withhold outstanding payments.

Is Venezuela still economically reliant on these platforms?

Posada: It has become less central to the platforms I studied. The companies have moved a lot of this work to countries with more workers who speak English, like India, the Philippines, and Kenya, as investment in large-language models has increased demand for language-based tasks. The broader economic context has also changed: Venezuela’s economy has rebounded from its pandemic-era low, although the recovery remains partial and uneven. Oil prices rose following the Russian invasion of Ukraine in 2022 and have risen again with the war involving Iran this year. The situation is very different from what it was six years ago.

What can be done to improve conditions for these workers, and is it all likely that it will be done?

Posada: It’s tough. It’s very easy for these companies to just move production around. Regulation should target the countries where major clients are based, particularly the United States. But I don’t see that happening anytime soon, unless Congress starts thinking about it after the midterms.